Module 17 Lesson 3: LlamaIndex security and data connectors

LlamaIndex is the specialist framework for Data. It is used to build knowledge bases (Module 10). Its security is all about Data Integrity.

1. Loader Security

LlamaIndex has hundreds of "Loaders" (e.g., PDFReader, SlackReader, GithubRepositoryReader).

• The Risk: Some loaders might execute code to "parse" a complex file. If you load a malicious .docx or .html file, it could exploit a vulnerability in the loader library itself.
• The Defense: Always use the Safest Reader possible. For example, use a plain-text loader instead of one that executes JavaScript to render a webpage.

---

2. Preventing "Context Window Overflow"

If a loader finds a 1GB text file, it will try to "Index" it.

• The Attack: Resource Exhaustion. An attacker uploads a massive "Zip Bomb" style text file to your data source. LlamaIndex will crash the server trying to turn it into embeddings.
• The Defense: Set strict Character Limits and Timeout Limits on every data connector.

---

3. Connector Token Security

To read your Slack or Google Drive, LlamaIndex needs an API Token.

• The Risk: Developers often put these tokens in their environment file but don't protect the server.
• The Defense: Use Role-based Service Accounts. Don't give LlamaIndex the "Admin" token for Slack; give it a "Viewer" token that can only see specific public channels.

---

4. Validating the "Index"

An Index in LlamaIndex can be saved to disk (e.g., as a .json file).

• The Attack: An attacker modifies the index.json file to change what facts the AI "retrieves." This is a Knowledge-base Hijack.
• The Defense: Use Checksums or Signatures to verify that your index files haven't been modified on the server.

---

Exercise: The Data Guardian

1. Why is "GitHub Repository Loader" a high-risk tool? (Hint: Think about what else is in a GitHub repo besides code).
2. You are building an AI for a law firm. Should you index the client's "Personal" email folder? Why/Why not?
3. How can you use "Metadata Filters" in LlamaIndex to implement the ACLs we learned in Module 10?
4. Research: What is the "LlamaCloud" security model and how does it handle private data storage?

---

Summary

LlamaIndex security is about What the AI knows. If the knowledge is poisoned, the AI is poisoned. By securing the connectors and loaders, you ensure that the AI's "Training" or "Retrieval" data remains pure and safe.

Next Lesson: The Secure Gateway: Middleware and proxy security for LLMs.